RTX 3090 Ti LLM benchmark
No signed RTX 3090 Ti submissions yet. Check the configuration guidance below, or contribute a reproducible measurement from your own machine.
Specs vendor specs, not measured
- VRAM
- 24 GB
- Memory bandwidth
- 1008 GB/s
- Launch date
- 2022-03-29
Sources: nvidia.com, nvidia.com, images.nvidia.com. Vendor-published figures, accessed 2026-10-14.
Plan memory before choosing a model
The catalog lists 24 GB of VRAM for this hardware entry. Check the capacity of your actual card. Model weights, context cache and runtime overhead must share the available memory. Capacity alone does not establish which model will fit or how fast it will run. Use the memory-fit calculator for your model and context as an estimate, then verify with the intended runtime. No signed RTX 3090 Ti measurement is available yet.
Contribute a signed RTX 3090 Ti benchmark
Run on YOUR hardware to populate this page: pipx install llm-speed && llm-speed bench
$ pipx install llm-speed && llm-speed bench
Community folklore on RTX 3090 Ti
9 unverified claims extracted from Reddit/HN comments. Lower trust than signed runs above; every row links to the source.
- communityconfidence 75%
30.00tok/s — Qwen-3 Coder on RTX 3090 Ti via ollama FP16
our signed data: RTX 3090 Ti · Qwen-3 Coder
“on: **FP16** * Hardware: **RTX 5090 + RTX 3090 Ti** * Task: code generation **Results:** * **llama.cpp:** \~52 tokens/sec * **Ollama:** \~30 tokens/sec Both runs use the same model weights and hardware. The gap is \~70% in favor of llama.cpp. Has anyone dug into why this happ…”
- communityconfidence 75%
52.00tok/s — Qwen-3 Coder on RTX 3090 Ti via llama.cpp FP16
our signed data: RTX 3090 Ti · Qwen-3 Coder
“**Qwen-3 Coder 32B** * Precision: **FP16** * Hardware: **RTX 5090 + RTX 3090 Ti** * Task: code generation **Results:** * **llama.cpp:** \~52 tokens/sec * **Ollama:** \~30 tokens/sec Both runs use the same model weights and hardware. The gap is \~70% in favor of llama.cpp. Has…”
- communityconfidence 75%
55.00tok/s — Mistral Nemo on RTX 3090 Ti via lm-studio FP16
our signed data: RTX 3090 Ti · Mistral Nemo
“s on both of the GPUs, so for example Mistral Nemo 13B instruct in FP16 and get 21.5 tokens/sec. Llama 3.1 Instruct Q8 15Gb results in 51 - 55 tokens per second. I also tried running Llama 3.1 Q5 KS, 49GB model which worked but was extremely slow due to some of the model going to…”
- communityconfidence 75%
21.50tok/s — Mistral Nemo on RTX 3090 Ti via lm-studio FP16
our signed data: RTX 3090 Ti · Mistral Nemo
“AI and LLM models, now I can completely offload latest models on both of the GPUs, so for example Mistral Nemo 13B instruct in FP16 and get 21.5 tokens/sec. Llama 3.1 Instruct Q8 15Gb results in 51 - 55 tokens per second. I also tried running Llama 3.1 Q5 KS, 49GB model which wor…”
- communityconfidence 60%
46.00tok/s — Llama 3 8B on RTX 3090 Ti via ollama
our signed data: RTX 3090 Ti · Llama 3 8B
“del locally. I have since tried both mlc-llm as well as ollama (based on llama.cpp). mlc-llm is slightly faster (\~51 tok/s) vs ollama (\~46 tok/s) for running the 16 bit unquantized version of Llama 3 8B on my RTX 3090 Ti. However, mlc-llm uses about 2GB of VRAM more than the …”
- communityconfidence 60%
51.00tok/s — Llama 3 8B on RTX 3090 Ti via ollama
our signed data: RTX 3090 Ti · Llama 3 8B
“s the way to run the model locally. I have since tried both mlc-llm as well as ollama (based on llama.cpp). mlc-llm is slightly faster (\~51 tok/s) vs ollama (\~46 tok/s) for running the 16 bit unquantized version of Llama 3 8B on my RTX 3090 Ti. However, mlc-llm uses about 2GB…”
- communityconfidence 60%
500.0tok/s — Llama 3 8B on RTX 3090 Ti via vllm
our signed data: RTX 3090 Ti · Llama 3 8B
“10 mins I can go through 24.9M prompt tokens and 1.69M generated tokens on 2x rtx 3090 Ti and Llama 3 8B W8A8 in vLLM 0.8.3 V1, so about 41,500 t/s input and 2800 t/s output, both at the same time. I went through like 8B+ tokens on single card this way before I put in the second …”
- communityconfidence 55%
3.50tok/s — on RTX 3090 Ti via llama.cpp EXL3
our signed data: RTX 3090 Ti
“bably won't be smart, but it will be stupid much faster since it will be all be in VRAM with llama cpp and IQ3_XSS GLM 4.6/4.7 I had about 3.5 t/s generation speed. 3200 DDR4 128GB and 2x rtx 3090 ti 24gb.”
- communityconfidence 50%
90.00tok/s — Qwen 3 Coder Next on RTX 3090 Ti
our signed data: RTX 3090 Ti · Qwen 3 Coder Next
“but doesn't work. **UPDATE:** Just for kicks, I tested the same prompt on **Qwen 3.5** **35B-A3B Q4 KXL UD** at **max context** and got **90 tok/sec**. :) However, I gave it 3 attempts like the others below, and while it loaded the GUI on output #3, the app didn't have the butto…”
Request a run on RTX 3090 Ti
No measured run yet for these pairs. Vote for the one you want; the most-requested runs go first. See all requests.
We never sell or share your email. Privacy